Luma AI, the San Francisco startup known for its video generation models, has raised a $900 million Series C led by HUMAIN, Saudi Arabia's sovereign AI company. As part of the deal, Luma will start running workloads on Saudi Arabia's Project Halo supercluster within weeks.
The money is the headline. The more interesting part is where the computing will happen.
Project Halo and why Luma wants it
Project Halo is a 2-gigawatt AI supercluster under construction in Saudi Arabia. Most large data centres run at 50 to 100 megawatts, so Halo is 20 to 40 times that size and among the largest compute installations in the world.
HUMAIN operates under Saudi Arabia's Public Investment Fund (PIF). It is building Halo as part of the Kingdom's aim to become the world's third-largest AI hub after the United States and China. Saudi Arabia has committed over $100 billion to AI-related infrastructure projects through 2030.
For Luma, the attraction is simple. Training frontier video generation and "world models" takes a great deal of compute. The company says its next systems will process video, audio, image and language data at "peta-scale", handling 1,000 to 10,000 times more information than current large language models. World models are systems that learn how the physical world behaves from video and other multimodal data.
Luma brings working technology. Its video models already power commercial products, and its Dream Machine attracted millions of users. CEO Amit Jain called HUMAIN "the perfect partner for this next stage in Luma AI's explosive trajectory". In practice, that means access to chips and compute that would be hard to get elsewhere.
An Arabic model and a Riyadh office
The partnership has a named deliverable. HUMAIN Create, described as "the world's first Saudi-built, Arabic-native foundation model", will be developed through the deal. It is meant to serve "enterprises and governments across MENA" and will be trained on Arabic and regional data, so it is a model built here rather than a licence for a Western one.
Luma is also opening a Riyadh office to support client engagement and regional development. That creates jobs for AI practitioners in the Kingdom. It also raises a fair question about whether the region has enough trained engineers to staff facilities of this scale.
The chips behind the deal
Luma expects to acquire up to 400,000 AI chips over the next four years through this partnership. Chips of this class have been in short supply for years, and many Silicon Valley companies struggle to secure allocations.
US export policy explains part of this. The Biden administration cleared major chip sales to Gulf allies, and the Trump administration has kept that approach, on the view that blocking allied access would push partners towards Chinese alternatives. For a startup like Luma, a Gulf partner can be a faster route to capacity than competing with hyperscalers for US supply.
Where the UAE fits
Saudi Arabia is one of several Gulf states investing in AI infrastructure. The UAE continues to invest through G42 and its partnerships with Microsoft and others. Qatar has positioned itself as a hub for AI conferences and collaboration, and Bahrain and Oman are developing more targeted strategies suited to their size.
The Saudi bet is on scale: that compute will be the main constraint on AI progress, and that whoever controls a lot of it gains influence. Compute alone will not settle that. Software skills, research talent, regulation and a working commercial market all matter too. Two gigawatts is still a serious start, and data centre builders across the region will watch how Halo handles power, cooling and day-to-day operations.
For a UAE organisation, the practical effect may arrive sooner than expected. Arabic-native models trained on regional data could become available on regional infrastructure, as an alternative to foreign cloud providers. If a vendor pitches you HUMAIN Create or a similar model, ask them three specific things in writing: where your prompts and documents are processed and stored, what Arabic data the model was trained on, and how it scores on Arabic tasks that look like your own work.
Sources: